SonarLatentOperationFilteredNoise

SonarLatentOperationFilteredNoise is a custom noise type that runs your noise through one or more LATENT_OPERATIONs before it becomes the noise you use. It's the bridge between the pack's two subsystems: the noise generators on one side, the operation machinery (scheduling, filtering, CFG-apply nodes) on the other. If you've ever wanted to take generated noise and then process it with an operation - clip its outliers, inject structure, reshape it - this is the node that lets you do that inside a noise chain.

How it works

custom_noise is your source chain. Then operation_1 through operation_5 - each optional - apply to the noise in sequence, in the order you plug them in. Attach a quantile filter and a pattern break, and your noise gets clipped then shuffled, in that order. Chain more than one and you're building a mini noise-processing pipeline out of reusable operations, which is the point: the same operations you use mid-sampling can now be used to shape the noise itself.

The inputs that matter

• custom_noise - the noise being filtered.
• operation_1 … operation_5 - the operations, applied in sequence. Order matters; swapping them changes the result.
• normalize_noise - whether the source noise is normalized initially, before any operation runs. Default false - useful if you want the operations to see the noise at its natural scale first.
• normalize - whether the final result is normalized to 1.0 strength. Default disabled. You usually want this on (or at least default) after operations have reshaped things, because operations can leave the noise at a weird magnitude.

Plus the standard factor and rescale chain controls. Output is SONAR_CUSTOM_NOISE, so the filtered noise flows into a Sonar sampler, NoisyLatentLike, or onward into the rest of the toybox.

A recipe that makes sense

A genuinely useful combo: SonarCustomNoise (say pyramid) → SonarLatentOperationFilteredNoise with a SonarLatentOperationQuantileFilter attached. The pyramid noise gives you multiscale structure, the quantile filter tames its long-tailed extremes, and the result is structured starting noise that doesn't blow out your early steps. Add normalize set to forced and you get a consistent magnitude on top. That's a full workflow in two nodes plus one operation.

The gotcha

If you attach operations and see no change, check normalize_noise and normalize before anything else - an operation that's normalizing your noise back to 1.0 on both ends can cancel out most visible difference. And remember operations run in plug order, so the classic mistake of "I added two filters and swapped them expecting the same result" will quietly bite you.

Install

ComfyUI Manager → search "ComfyUI-sonar" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/blepping/ComfyUI-sonar

No models, no extra deps. Usual pack caveats: inputs shift between versions, noise seeds can differ across GPU models, and the author's standing disclaimer that the noise types are best-effort implementations applies upstream of this node.